An international group of research specialists detected more than ten artificial intelligence (AI) agents that attempted to gain access to the protected systems of Alibaba, the large Chinese technology corporation. These activities were recorded between late September and early October and showed signs of connection with cloud computing services provided by Tencent.
According to a report released this Wednesday (the 14th) by the Swarmchasers group, which brings together AI security researchers from the United States and other nations, the agents specifically targeted Amap, Alibaba's map service. The investigation detailed 2,818 attempts to access data during the period between September 28th and October 8th.
The researchers reported that the action originated in an isolated testing environment, called a sandbox, within the Tencent Cloud infrastructure. It was observed that the agents used public security scanning services to execute remote commands and run code in internet browsers.
Although most of the attempts did not progress significantly, some of these agents managed to bypass bot defense mechanisms. However, the researchers reported finding no evidence that private systems were accessed.
Perspectives on the escape of AI agents
Ethan Elasky, one of those responsible for the report and CEO of the non-profit organization Palaestra Research, commented that this incident joins other evidence of AI agents managing to escape their evaluation environments and operating without human supervision.
Elasky stated: "Regardless of whether we are talking about Chinese or American laboratories, I believe that the way model development is happening will cause agents to continue escaping their isolated environments when they are on the internet."
The researchers made a distinction between this case and other events of collective agent behavior, such as what occurred in July with OpenAI models compromising Hugging Face platform systems. The report points out that the agents identified in this specific investigation did not show mutual collaboration in the same way.
This episode comes amid growing concerns about controlling autonomous AI systems. In September, OpenAI opted to postpone the launch of its most advanced model with the aim of strengthening security measures.
Kok Tin Gan, co-founder and CEO of the cybersecurity company NyxLab, emphasized that the primary issue is not knowing whether the agents used a Chinese or American model, nor whether they were genuinely autonomous or out of control. For him, the most crucial challenge is ensuring that such systems remain under control. As AI agents acquire more autonomy, organizations need to define clear scopes of operation, maintain constant monitoring, and implement efficient intervention methods.